Adaptation algorithms for HTTP-based video streaming
Miller, Konstantin · DepositOnce · 2016
Ever since the invention of the cinematography, there has been a growing demand for high-quality video content. Recently, the broad availability of high-speed wireless Internet access, complemented by the pervasiveness of mobile, computationally powerful devices with high-resolution screens, have made the video delivery over the open Internet the technology of choice for both video on demand and live streaming services. Due to the best-effort nature of the Internet, however, ensuring a high quality of experience is challenging. A state-of-the-art approach to address this challenge is adaptive streaming, designed to continuously adjust the characteristics of the streamed media to dynamically varying network conditions, leading to a smoother viewing experience with less playback interruptions and a more efficient utilization of the available network resources. Despite the ongoing efforts, however, recent studies suggest that the challenge has not yet been successfully resolved. One of the open issues is the design of efficient adaptation algorithms, that are among the primary factors determining the overall performance of a streaming service. In this thesis, I present several contributions to this area of research, that are outlined in the following. In order to cope with the wireless traffic increase expected over the next years, it will be necessary to increase the density of the deployed wireless infrastructure. In my first contribution, I focus on a simultaneous delivery of a large number of unicast video on demand streams in a dense wireless network. I jointly consider the problem of wireless transmission scheduling and video quality selection, and develop a distributed approach based on control theory. The conducted performance evaluation shows that the presented approach is able to serve an up to twice as large number of users completely without interruptions, as compared to a baseline approach. Simultaneously, it allows to reduce the number of quality transitions by up to 50%, without reducing the average video quality. In addition, the unfairness among the individual streaming sessions is reduced by up to a factor of 4.Even though the majority of the video content being streamed over the Internet is video on demand, the amount of live streaming is growing rapidly. In my second contribution, I focus on a particularly challenging use case of low-delay live streaming. I develop a novel adaptation algorithm that is leveraging throughput predictions to provide a high quality of experience over wireless links, with a latency bound on the order of a few seconds. It heuristically maximizes the average video quality at an operating point defined by the live latency, amount of playback interruptions, and number of quality transitions. A comparative evaluation reveals that at the individual operating points, the developed algorithm provides an average video quality which is by up to a factor of 3 higher than the quality achieved by the baseline approach. Furthermore, it is able to reach a broader range of operating points, and can thus be more flexibly adapted to the user profile and service provider requirements. In my third contribution, I develop a universal adaptation algorithm for video on demand, that can operate over a broad range of network conditions, and that has a flexible configuration that can be adjusted to the particular service and user requirements. It uses the playback buffer level information and the past throughput information to meet its adaptation decisions. It does not rely on a cooperation with the network nor on cross-layer information, and is therefore suitable for a standalone deployment in any network environment, and on a broad range of platforms. Moreover, it minimizes the start-up delay, which is particularly important for services, where users tend to frequently start new video sessions. I evaluate the approach against a baseline and against an omniscient client that computes optimal adaptation trajectories by solving a series of optimization problems. The evaluation reveals that the proposed algorithm allows to efficiently avoid playback interruptions, provides a smooth viewing experience by avoiding excessive video quality fluctuations, achieves a high level of network resource utilization, and provides a fair resource allocation in a multi-user environment. In particular, in the network environment used for the evaluation, the developed algorithm achieves an average video bit rate which is by up to 35% higher than that of the baseline approach, and within up to 85% of the optimum, with an up to an order of magnitude smaller total duration of interruptions. It is worth mentioning that the omniscient client developed in the course of this work can not only serve as a reliable benchmark for streaming clients but also allows to evaluate the influence of various media and network properties on the achievable streaming performance. Last but not least, based on my experience with implementing streaming client prototypes and simulation models, I develop a streaming client architecture that is modular, extendible, and platform-independent, and efficiently supports distributed operation of the individual functional blocks.